Add new variance update option and unittests for intelligibility

webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc

 /*
  *  Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
  *  in the file PATENTS.  All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
 //
 //  Implements core class for intelligibility enhancer.
 //
 //  Details of the model and algorithm can be found in the original paper:
 //  http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6882788
 //
 
 #include "webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.h"
 
 #include <cmath>
 #include <cstdlib>
 
 #include <algorithm>
 #include <numeric>
 
 #include "webrtc/base/checks.h"
 #include "webrtc/common_audio/vad/include/webrtc_vad.h"
 #include "webrtc/common_audio/window_generator.h"
 
+namespace {

[Andrew MacDonald, 2015/07/02 02:46:48]

Not a big deal, but normally this is still wrapped in the webrtc namespace.

[ekm, 2015/07/07 21:57:02]

Done.

+
+const int kErbResolution = 2;
+const int kWindowSizeMs = 2;
+const int kChunkSizeMs = 10;  // Size provided by APM.
+const float kClipFreq = 200.0f;
+const float kConfigRho = 0.02f;  // Default production and interpretation SNR.
+const float kKbdAlpha = 1.5f;
+const float kLambdaBot = -1.0;      // Extreme values in bisection
+const float kLambdaTop = -10e-18f;  // search for lamda.
+
+}  // namespace
+
+namespace webrtc {
+
 using std::complex;
 using std::max;
 using std::min;
-
-namespace webrtc {
-
-const int IntelligibilityEnhancer::kErbResolution = 2;
-const int IntelligibilityEnhancer::kWindowSizeMs = 2;
-const int IntelligibilityEnhancer::kChunkSizeMs = 10;  // Size provided by APM.
-const int IntelligibilityEnhancer::kAnalyzeRate = 800;
-const int IntelligibilityEnhancer::kVarianceRate = 2;
-const float IntelligibilityEnhancer::kClipFreq = 200.0f;
-const float IntelligibilityEnhancer::kConfigRho = 0.02f;
-const float IntelligibilityEnhancer::kKbdAlpha = 1.5f;
-
-// To disable gain update smoothing, set gain limit to be VERY high.
-// TODO(ekmeyerson): Add option to disable gain smoothing altogether
-// to avoid the extra computation.
-const float IntelligibilityEnhancer::kGainChangeLimit = 0.0125f;
-
 using VarianceType = intelligibility::VarianceArray::StepType;
 
 IntelligibilityEnhancer::TransformCallback::TransformCallback(
     IntelligibilityEnhancer* parent,
     IntelligibilityEnhancer::AudioSource source)
     : parent_(parent), source_(source) {
 }
 
 void IntelligibilityEnhancer::TransformCallback::ProcessAudioBlock(
     const complex<float>* const* in_block,
@@ skipping 136 matching lines @@
       ProcessClearBlock(in_block, out_block);
       break;
     case kCaptureStream:
       ProcessNoiseBlock(in_block, out_block);
       break;
   }
 }
 
 void IntelligibilityEnhancer::ProcessClearBlock(const complex<float>* in_block,
                                                 complex<float>* out_block) {
-  float power_target;
-
   if (block_count_ < 2) {
     memset(out_block, 0, freqs_ * sizeof(*out_block));
     ++block_count_;
     return;
   }
 
   // For now, always assumes enhancement is necessary.
   // TODO(ekmeyerson): Change to only enhance if necessary,
   // based on experiments with different cutoffs.
   if (has_voice_low_ || true) {
     clear_variance_.Step(in_block, false);
-    power_target = std::accumulate(clear_variance_.variance(),
-                                   clear_variance_.variance() + freqs_, 0.0f);
+    float power_target = std::accumulate(

[Andrew MacDonald, 2015/07/02 02:46:48]

const

[ekm, 2015/07/07 21:57:02]

Done.

+        clear_variance_.variance(), clear_variance_.variance() + freqs_, 0.0f);
 
     if (block_count_ % analysis_rate_ == analysis_rate_ - 1) {
       AnalyzeClearBlock(power_target);
       ++analysis_step_;
       if (analysis_step_ == variance_rate_) {
         analysis_step_ = 0;
         clear_variance_.Clear();
         noise_variance_.Clear();
       }
     }
     ++block_count_;
   }
 
   /* efidata(n,:) = sqrt(b(n)) * fidata(n,:) */
   gain_applier_.Apply(in_block, out_block);
 }
 
 void IntelligibilityEnhancer::AnalyzeClearBlock(float power_target) {
   FilterVariance(clear_variance_.variance(), filtered_clear_var_.get());
   FilterVariance(noise_variance_.variance(), filtered_noise_var_.get());
 
-  // Bisection search for optimal |lambda|
+  SolveForGainsGivenLambda(kLambdaTop, start_freq_, gains_eq_.get());
+  float power_top =

[Andrew MacDonald, 2015/07/02 02:46:48]

const

[ekm, 2015/07/07 21:57:02]

Done.

+      DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
+  SolveForGainsGivenLambda(kLambdaBot, start_freq_, gains_eq_.get());
+  float power_bot =

[Andrew MacDonald, 2015/07/02 02:46:48]

const

[ekm, 2015/07/07 21:57:02]

Done.

+      DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
+  if (power_target >= power_bot && power_target <= power_top) {
+    SolveForLambda(power_target, power_bot, power_top);
+    UpdateErbGains();
+  }  // Else experiencing variance underflow, so do nothing.
+}
 
-  float lambda_bot = -1.0f, lambda_top = -10e-18f, lambda;
-  float power_bot, power_top, power;
-  SolveForGainsGivenLambda(lambda_top, start_freq_, gains_eq_.get());
-  power_top =
-      DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
-  SolveForGainsGivenLambda(lambda_bot, start_freq_, gains_eq_.get());
-  power_bot =
-      DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
-  DCHECK(power_target >= power_bot && power_target <= power_top);
-
+void IntelligibilityEnhancer::SolveForLambda(float power_target,
+                                             float power_bot,
+                                             float power_top) {
+  float lambda_bot = kLambdaBot;
+  float lambda_top = kLambdaTop;
   float power_ratio = 2.0f;  // Ratio of achieved power to target power.
   const float kConvergeThresh = 0.001f;  // TODO(ekmeyerson): Find best values
   const int kMaxIters = 100;             // for these, based on experiments.
   int iters = 0;
   while (fabs(power_ratio - 1.0f) > kConvergeThresh && iters <= kMaxIters) {
-    lambda = lambda_bot + (lambda_top - lambda_bot) / 2.0f;
+    float lambda = lambda_bot + (lambda_top - lambda_bot) / 2.0f;

[Andrew MacDonald, 2015/07/02 02:46:48]

const

[ekm, 2015/07/07 21:57:02]

Done.

     SolveForGainsGivenLambda(lambda, start_freq_, gains_eq_.get());
-    power = DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
+    float power =

[Andrew MacDonald, 2015/07/02 02:46:47]

const

[ekm, 2015/07/07 21:57:02]

Done.

+        DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
     if (power < power_target) {
       lambda_bot = lambda;
     } else {
       lambda_top = lambda;
     }
     power_ratio = fabs(power / power_target);
     ++iters;
   }
+}
 
+void IntelligibilityEnhancer::UpdateErbGains() {
   // (ERB gain) = filterbank' * (freq gain)
   float* gains = gain_applier_.target();
   for (int i = 0; i < freqs_; ++i) {
     gains[i] = 0.0f;
     for (int j = 0; j < bank_size_; ++j) {
       gains[i] = fmaf(filter_bank_[j][i], gains_eq_[j], gains[i]);
     }
   }
 }
 
@@ skipping 116 matching lines @@
                                           int length) {
   float ret = 0.0f;
 
   for (int i = 0; i < length; ++i) {
     ret = fmaf(a[i], b[i], ret);
   }
   return ret;
 }
 
 }  // namespace webrtc